List of Papers Browse by Subject Areas Author List
Abstract
Uncertainty arising from inter-observer variability in medical image segmentation plays an important role in developing treatment plans. Research in this area is inhibited by the lack of multiple annotations for large-scale medical datasets, especially for volumetric data, which suffers from additional scaling and computational complexity challenges. Flow matching has emerged as a powerful framework for generative modelling and has also been demonstrated to maintain strong performance when working with latent representations of images. In this work, we introduce a latent-to-latent flow technique for stochastic segmentation of medical volumes via encoded representations of both the image and label space. We evaluate our method on two challenging applications covering delineation uncertainty for radiotherapy planning and multiple organ structure segmentation, improving efficiency up to 14x compared with full resolution models while maintaining clinically relevant performance. Code available at: https://github.com/biomedia-mira/L2L-Flow.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3260_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/biomedia-mira/L2L-Flow
Link to the Dataset(s)
N/A
BibTex
@InProceedings{TodOma_LatenttoLatent_MICCAI2026,
author = { Todd, Omar AND Kim, Sooha AND Mehta, Raghav AND Mackay, Katherine AND Bernstein, David AND Taylor, Alexandra AND De Sousa Ribeiro, Fabio AND Glocker, Ben},
title = { { Latent-to-Latent Flow for Volumetric Stochastic Segmentation } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16884},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The paper addresses a real bottleneck in volumetric stochastic segmentation, namely the computational cost of full-resolution generative modeling.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
This is a solid paper on an important problem: making stochastic segmentation practical for 3D volumes. The paper is technically clear, the proposed latent-to-latent formulation is sensible, and the efficiency motivation is well justified. I especially liked that the paper does not oversell the limitations of existing metrics and includes qualitative clinical context. The main caveat is that the performance story is mixed: the method is attractive primarily as an efficiency tradeoff rather than as a clear accuracy improvement over the strongest full-resolution baseline.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
The strongest full-resolution baseline still wins on the main quality metric. On the radiotherapy dataset, Flow-SSN with the shifted schedule achieves the best GED, while L2L-Flow is presented as competitive but not superior. On CURVAS, the gap is more noticeable, especially for the pancreas, and the paper needs a higher-resolution 2x latent variant to recover part of that gap. The paper should frame the contribution more explicitly in those terms.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not provide sufficient information for reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(5) Accept — should be accepted, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
See strength.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #2
- Please describe the contribution of the paper
This paper identifies performance degradation of Flow-SSNs on high-resolution volumetric medical data, and introduces a simple time-shifted noise schedule resulting in significantly improved stability and performance at scale.
It also introduces the Latent-to-Latent Flow (L2L-Flow), an efficient and scalable volumetric stochastic segmentation approach that enables up to 14× faster inference compared to Flow-SSNs, whilst retaining competitive performance.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
The paper addresses a real bottleneck: stochastic segmentation in 3D is costly, and uncertainty from inter-observer variability matters clinically. That motivation is well stated and relevant.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
[1] The novelty feels somewhat incremental. The paper combines known ingredients: flow matching / Flow-SSN, latent generative modeling, label autoencoding, a shifted schedule inspired by prior latent flow work. The combination is useful, but the conceptual novelty may be viewed as moderate rather than major. The strongest contribution is likely engineering and empirical scaling, not a fundamentally new formulation.
[2] Baseline fairness could be discussed more carefully. The models are compared at fixed ROI size and similar training setup, but some baselines may not be fully optimized for 3D or for the exact computational budget used here. The paper notes challenges with MoSE and SSN in 3D, but the discussion could better separate “methodological weakness” from “insufficient adaptation/tuning.
[3] Flow matching may increase the trainning time and the trainning time should be reported to analysis the effiency.
[4] How does it compared with diffusion based appraoch? Because diffusion model are also an generative model.
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The novelty is a bit limited; more experiments should be compared with more baselines, as well as the diffusion model method. In addition, a abalation study of efficiency should be reported.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Reject
- [Post rebuttal] Please justify your final decision from above.
The novelty seems quite limited. In addition, this paper lacks a diffusion model baseline.
Review #3
- Please describe the contribution of the paper
Uncertainty arising from inter-observer variability in medical image segmentation plays an important role in treatment planning. In this work, the authors introduce a latent-to-latent flow technique for stochastic segmentation by leveraging encoded representations of both the image and label spaces. The method is evaluated on two applications and demonstrates improved efficiency while maintaining competitive performance.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
1.The paper proposes a modified shift schedule that significantly improves training stability and performance of Flow-SSNs in volumetric settings.
2.The method builds on Flow-SSNs and introduces a novel approach that learns a flow transport from a latent, volume-conditional prior to a latent label distribution.
3.The results across two applications demonstrate strong clinical potential.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
1.Based on the reported results, L2L-Flow appears to achieve performance similar to Flow-SSNs, aside from improvements in inference time. The paper would benefit from a clearer discussion of why faster inference is clinically important and how it impacts real-world workflows.
2.Uncertainty modeling is presented as a key motivation, but the paper provides limited analysis of uncertainty. More comprehensive evaluation and discussion are needed, particularly in relation to the two clinical application scenarios.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not provide sufficient information for reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The proposed method builds on Flow-SSNs, but the reported performance gains are limited, which raises some concerns regarding the level of novelty and practical improvement over existing approaches. While efficiency improvements are noted, their clinical impact is not sufficiently discussed. In addition, uncertainty modeling is presented as a key motivation of the work, yet the paper lacks a thorough uncertainty analysis and does not clearly demonstrate how this aspect contributes to the proposed clinical applications.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #4
- Please describe the contribution of the paper
This paper proposes a latent-to-latent (L2L) flow-based stochastic segmentation framework as an extension of Flow Stochastic Segmentation Networks (SSNs). Instead of modeling stochasticity in pixel space, the authors perform uncertainty modeling in latent space using a dedicated mask encoder–decoder and a latent flow prior. The method aims to address computational and scalability limitations of pixel-space flows, particularly for volumetric medical image segmentation, and is evaluated on multiple datasets including a radiotherapy delineation task.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
- The work addresses a relevant and well-known limitation of flow-based SSNs, namely their reliance on pixel-space transformations, which are computationally expensive for volumetric data.
- The proposed latent-to-latent formulation is conceptually sound and well motivated.
- The application to uncertainty-aware segmentation in radiotherapy is timely and practically relevant.
- Reported inference speed improvements indicate potential for clinical usability.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
1.Insufficient architectural detail o The paper does not provide sufficient architectural information for the mask encoder–decoder (e.g., number of channels, latent dimensionality, depth). o Similarly, architectural details of the latent flow/prior model are underspecified. o This lack of detail makes it difficult to assess the expressive capacity of the latent space, compare with prior SSN implementations, or reproduce the method. 2.Ambiguity in latent prior design o It is unclear whether the latent flow model is architecturally symmetric for image and label latents or whether different capacities are used. o Explicit description of the flow architecture would improve clarity and reproducibility. 3.Downsampling factor choices (×2 vs ×4) o Different downsampling factors are used across datasets, but no explicit justification is provided. o An ablation study comparing ×2 and ×4 downsampling—especially added to the quantitative results table—would strengthen the analysis and help clarify the trade-offs between accuracy and efficiency. 4.Radiotherapy uncertainty analysis clarity o In the uncertainty visualization and contour analysis, it is unclear whether the “SSN contours” correspond to pixel-space SSN, flow-based SSN, or the proposed L2L Flow model. o This ambiguity makes it difficult to interpret the comparative uncertainty results. 5.Inference speed vs. performance trade-off o The reported ~14× speedup appears to correspond to the ×4 downsampled model, whereas the paper notes that the ×2 model achieves better segmentation performance. o It would be beneficial to include qualitative results for the ×2 model, as well as a discussion of its runtime trade-offs. 6.Consistency of experimental reporting o For robustness and transparency, results for both ×2 and ×4 configurations should be reported consistently across tables, quantitative metrics, and qualitative examples. 7.Cross-validation protocol details o While a 5-fold cross-validation strategy is mentioned, the exact training/validation splits are not clearly described. o The reference to “16 samples” is ambiguous: it is unclear whether these correspond to validation cases, test cases, or randomly selected samples. o Clarifying these points would improve experimental rigor. 8.Lack of ablation on latent generation design o No ablation study is provided to justify the use of a dedicated mask encoder–decoder over simpler latent representations obtained via standard downsampling/upsampling. o Such an analysis would strengthen the methodological justification of the proposed pipeline.
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not provide sufficient information for reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper presents a promising and meaningful extension of flow-based stochastic segmentation to latent space, with clear motivation and relevance to volumetric medical imaging. However, the lack of architectural clarity, limited ablation studies, and some inconsistencies in experimental reporting currently weaken the paper’s technical transparency and reproducibility. Addressing these concerns would significantly strengthen the submission.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
After considering the authors’ responses, I would like to maintain my weak accept recommendation. The rebuttal does a good job of clarifying the intended contributions, particularly the positioning of the latent-to-latent (L2L) flow framework and its role in enabling efficient 3D stochastic segmentation. I also appreciate the additional context provided around clinical relevance and uncertainty evaluation, especially the discussion involving inter-observer variation, which I found meaningful. That said, some of my earlier concerns are still only partially addressed. In particular, important architectural details of the mask encoder–decoder and the latent flow design remain unclear. While the authors mention that code will be released, having these details in the paper is important for assessing the method’s capacity and reproducibility at review time.Similarly, the discussion around downsampling choices (×2 vs ×4) would benefit from more complete and consistent quantitative comparisons. I also still feel that an ablation justifying the use of the dedicated latent generation module would strengthen the paper. Overall, I believe this work explores a promising and relevant direction, particularly in tackling scalability challenges of flow-based SSNs and improving efficiency for uncertainty-aware segmentation. However, clearer presentation and more complete experimental detail would further strengthen the contribution.
Author Feedback
We thank the reviewers for their valuable feedback and address the main concerns below
Novelty/Clinical Impact (R2, R4, R1): We politely push back on remarks about limited novelty by reiterating our two main technical contributions. Firstly, we identify a previously unknown issue with Flow SSNs, limiting direct application to 3D. We resolve this by proposing a time-shifted noise schedule tailored to the loss. The resulting method is not just a baseline but sets a new state-of-the-art in 3D stochastic segmentation. Our proposed solution achieves the strongest performance on two challenging segmentation tasks. Secondly, our L2L-Flow addresses efficiency (up to 14x faster inference) and offers a highly competitive alternative to our voxel-space Flow-SSN, with a controllable efficiency/performance trade-off. Importantly, L2L-Flow is tailored for stochastic segmentation, employing a two-sided interpolant between an image prior and a latent label distribution. This differs from standard latent flows where one side is a fixed noise distribution. L2L-Flow is also more easily applicable to different 3D problems compared to Flow-SSN, as noise schedule tuning is no longer required for strong performance. The efficiency benefits of L2L-Flow enable novel applications such as uncertainty-guided adaptive radiotherapy planning. Efficient contour generation allows real-time updates to treatment plans and reduces the likelihood of inter‑fraction anatomical changes between imaging and treatment delivery. As suggested by R1, we will make the framing of our contributions clearer in the final version.
Uncertainty Evaluation (R4): Our quantitative evaluation includes standard metrics such as GED, but also went further by evaluating clinical utility through expert review. This qualitative evaluation revealed key limitations of these metrics, where instances of contours with strong GED performance didn’t show clinically meaningful variation. This was especially true of low-rank SSNs as discussed in Section 4.3.We therefore performed further analysis, adding clinical context to our results (noted as a strength by R1), which is largely missing in other works. Fig. 3 evaluates contour variation of our models in relation to clinically meaningful inter-observer variation (IOV). Alongside GED, this enabled our clinicians to assess the usefulness of predicted contours. Importantly, our proposed flow-based methods exhibited the strongest alignment with expert IOV.
Baseline Comparisons (R2): We did not compare to diffusion models as rectified flows are from the same model family, and Flow-SSN was already shown to perform better than CCDM [28]. All baselines were faithful 3D adaptations of their 2D implementations and training procedure to ensure fairness. Baseline efficiency comparisons focused on inference time due to its importance in downstream clinical use. Training efficiency was similar across models. As demonstrated by our time-shifted 3D Flow-SSN, scaling and optimising 2D methods for 3D is non-trivial. We will make the nuance surrounding these technical challenges clearer in the final version.
Ablations (R2, R5): In the radiotherapy experiments, 4x downsampling was already sufficient to achieve comparable performance to our time-shifted 3D Flow-SSN. An efficiency ablation study was performed on CURVAS in Section 4.4.Due to the small size of the pancreas, we performed a further experiment at 2x downsampling to quantify the speed vs performance tradeoff, as shown in Fig. 4 & 5.A strength of L2L-Flow is its flexibility at adapting the latent capacity depending on data complexity. Scales such as 8x have potential application, and can be explored in future work. We will make the rationale behind our ablation choices clearer in the final version
Architecture Details (R5): We will release the code covering full implementation details with the publication. The implementation for the rectified flow was adapted from the open-source Flow-SSN codebase.
Meta-Review
Meta-review #1
- Your recommendation
Invite for Rebuttal
- Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.
Mixed reviews, but all reviewers acknowledge the potential of the contributions and established improvements. I would propose to give the authors the opportunity to address the raised concerns in a rebuttal.
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
Only one out of four reviewers submitted a post-rebuttal score. The original reviews were mixed 5-3-3-4, and in addition to the 3 that remained a reject, I got an email from the 4 that proposed to accept, as her/his concerns were mostly addressed and mentioning that “…this work explores a promising and relevant direction, particularly in tackling scalability challenges of flow-based SSNs and improving efficiency for uncertainty-aware segmentation. However, clearer presentation and more complete experimental detail would further strengthen the contribution.”. Still very much a borderline case.
Meta-review #2
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
This work is interesting due to its focus on a stochastic, uncertainty-aware segmentation attempt, which is shown to be made very efficient. Reviewer concerns regarding novelty and failed comparison to diffusion baselines were convincingly discussed. I think this is relevant work that should be discussed at the conference.
Meta-review #3
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Reject
- Please justify your recommendation.
After the rebuttal, R2 maintained the Reject recommendation, R5 maintained the Accept recommendation, and R1 and R4 did not provide updated reviews. Based on the authors’ rebuttal and the reviewers’ comments, the paper is not recommended for acceptance. Although the paper demonstrates some strengths, including addressing a practical problem and achieving promising improvements in inference speed, it also exhibits several fundamental limitations, including insufficient clarification of the performance–inference speed trade-off and its clinical utility, as well as insufficient uncertainty analysis. The authors are encouraged to address the reviewers’ comments when revising the paper for submission elsewhere.
